AI Frontier Daily

OpenAI announced three major scientific discoveries across mathematics, computer science, and physics, including resolving local smoothing conjectures to establish the Kakeya conjecture, completing half of the rigorous proof for Yang-Mills theory, and disproving the Exponential Time Hypothesis (ETH). These achievements mark a breakthrough in AI's formal reasoning capabilities for extremely abstract problems.

To address academic trust issues regarding raw AI-generated research data, OpenAI introduced external experts as "shepherds" for verification. Computer science and physics results undergo expert-guided review, while mathematical outputs are suggested for independent interpretation by researchers. This strategy aims to convert raw AI logic into community-accepted scientific knowledge through three distinct collaboration paths.

OpenAI released a library of mathematical manuscripts generated by internal models, including papers claiming to disprove the Exponential Time Hypothesis (ETH). The repository contains 721 manuscripts at various validation stages; some results have not yet passed Lean formal proofs and are considered raw research data rather than final academic outcomes.

OpenAI claims AI overturned ETH, a cornerstone of complexity theory, but did not prove P=NP. AI is also alleged to have proven the existence part of Yang-Mills theory, though the mass gap remains incomplete and details are undisclosed. All core claims are under independent external verification without traditional peer review, sparking debate on AI's role in scientific contribution.

Yanshan Tech, HKUST, and CMU jointly released and open-sourced the EdgeRock-4B model, reducing KV Cache by 99.1% via a fixed memory occupation mechanism. With constant VRAM usage at 256K context, single-card throughput reaches 6.42 times the source model while maintaining high reasoning accuracy, significantly lowering deployment costs for long-text and Agent scenarios.

KV Cache usage remains constant (72MiB) from 4K to 256K context length, a 99.1% reduction compared to traditional models. A single 80GB GPU supports 512 concurrent long-reasoning streams with an output throughput of 12,124 tokens/s, 6.42 times the source model. It achieves an average pass rate of 82.9% on AIME 2024-2026 benchmarks, far exceeding similar bounded-state models.

AI unicorn Anthropic formally filed its S-1 document with the SEC, planning to list on Nasdaq under code "ANT" with a target valuation of ~$2 trillion and aiming to raise up to $100 billion, potentially becoming the largest IPO in history. The prospectus reveals strong commercial momentum, with H1 2026 revenue reaching $16.3 billion, a 13x year-over-year increase, and positive operating profit in Q2.

Despite a net loss of $80.7 billion, sources attribute this primarily to non-cash accounting losses from fair value changes in Amazon and Google convertible bonds, with actual operational losses around $700 million. Key drivers include improved compute cost efficiency, dropping from 250% to 48% of revenue. The company disclosed long-term compute procurement contracts exceeding $500 billion and adopted a dual-class share structure to ensure founder control.

Meta announced a financing plan of up to $95 billion to support core businesses and emerging AI fields. The plan includes forward stock sales, convertible preferred shares, and potential common stock issuance, with projected 2026 capital expenditures of $135-145 billion to sustain continuous AI investment and return growth.

The financing structure introduces non-voting Class C shares, ensuring CEO Zuckerberg maintains absolute control despite significant fundraising. Chip giants NVIDIA, AMD, and Broadcom expressed interest in purchasing up to $5 billion worth of Meta Class A common stock, indicating deep supply chain integration and partner confidence in Meta's AI strategy.

Microsoft recently offered voluntary buyout packages to approximately 60,000 employees in the US, UK, Canada, and India, representing nearly 27% of its global full-time workforce. This move aims to optimize human resources through non-coercive means, addressing organizational adjustments and cost pressures in the generative AI era, marking a shift from passive layoffs to active structural restructuring among tech giants.

Sources indicate the plan covers core product teams like Copilot, CoreAI, Windows + Devices, and Xbox, explicitly excluding sales, marketing, quantum computing, and security departments. Eligible employees must decide by early December, choosing between a compensation package including severance, continued stock vesting, and healthcare for year-end departure, or declining the offer to retain their current positions.

Tsinghua University and NUS jointly proposed ParallelWorld, introducing test-time computation into high-level planning for embodied agents. By simulating multiple candidate action paths and comparing future observations, robots can predict consequences before execution, significantly boosting task success rates from 56% to 83%, offering a new paradigm for decision optimization in embodied intelligence.

ParallelWorld introduces rehearsal and verification of candidate futures between high-level planning and action execution; it uses Test-Time Scaling to compare subsequent observations and task progress across different action paths; experiments show this method increases embodied agent task success rates from 56% to 83%.

Google launched Android Bench 2.0, adding Long-Horizon Task (LHT) evaluation, agent-based closed-loop testing, and continuous scoring mechanisms to address challenges in assessing complex multi-step development tasks. The benchmark indicates AI performs better in writing new code than refactoring existing code, with Claude Opus 5.5 currently leading with a 32% pass rate.

It introduces LHTs covering engineering practices requiring days, such as dependency upgrades and building from scratch; adopts continuous scoring instead of binary judgments to improve recognition precision for partially successful behaviors; data shows AI excels more in generating new code than refactoring old code, with Claude Opus 5.5 temporarily ranking first.